1 INTRODUCTION

Hepatocellular carcinoma (HCC) is a prevalent cancer globally, ranking fifth in incidence and third in cancer-related deaths.¹ Radical surgery remains the primary method for curing HCC due to strict transplant conditions and a limited supply of liver donors for transplantation.² However, HCC has a high recurrence rate post-surgery, with rates ≥10% annually and reaching 30%–50% after 2 years.^{3, 4} Thus, identifying high-risk factors for postoperative recurrence of HCC and establishing a stable and effective predictive model is of utmost importance for the treatment and prognosis of patients.

As an essential component of standard management for patients with HCC, traditional imaging assessment relies heavily on qualitative features and lacks the ability to identify tumor heterogeneity.^{5, 6} Despite improvements in imaging technology, challenges persist in accurately assessing and monitoring tumors.⁷ In recent years, radiomics has been widely used for predicting tumor treatment efficacy, recurrence, and survival through quantitative analysis of high-throughput images, transforming the qualitative and subjective assessment of conventional imaging into quantifiable and reproducible standardized image information.^8-10 Studies have demonstrated the potential value of radiomics in high-throughput quantitative analysis of HCC images to effectively assist in the personalized management of high-risk patients for tumor recurrence and to characterize and screen patients at high risk of recurrence.^11-13 Although radiomics has shown promising results in predicting tumor recurrence, it is limited by potential hazards, such as overfitting and limited applicability. Therefore, further research is needed to explore and validate the reliability and validity of radiomics.

The wealth of information on pathological features in postoperative tissues provides essential information for clinical diagnosis, treatment, and prognosis.^{14, 15} However, the variability in evaluation between pathologists and the semi-quantitative nature of the scoring system makes it difficult to avoid the problems of reproducibility and poor subjectivity.^{16, 17} With the development of digital pathology, the foundation has been laid for rapid identification and accurate quantification of pathological features in pathomics.¹⁸ Recent studies have shown that pathomics and radiomics can effectively predict tumor recurrence, metastasis risk, and survival, providing an essential reference basis and guidance for clinical treatment and management.¹⁹ It should be noted that pathomics is different from radiomics in that the former provides deeper microscopic information about tumor cells and subcellular structures. The latter reflects the intensity distribution and spatial relationship of tumor tissues. Therefore, combining pathomics and radiomics can comprehensively capture tumors' macroscopic and microscopic structural features.^{20, 21} In recent studies, prognostic models that combine radiomics and pathomics features have demonstrated excellent predictive performance in various cancers, such as colorectal and glioblastoma multiforme.^{20, 22} However, in previous studies, explorations on predicting the risk of postoperative recurrence in patients with HCC have mainly focused on predictions based on preoperative computed tomography (CT)/ magnetic resonance imaging (MRI) radiomics and combined clinical features.^{23, 24} No studies have used radiomics and pathomics features to predict the risk of postoperative recurrence in patients with HCC.

Therefore, this study aimed to develop and validate a multimodal model using preoperative clinical, radiomics, and pathomics features to predict the risk of postoperative early recurrence in patients with HCC.

2 METHODS

2.1 Study population

This study aimed to develop a multimodal clinical-radiomic-pathomic model to predict postoperative early recurrence risk in patients with HCC who underwent radical resection. To achieve this goal, we conducted a retrospective analysis of data from patients who underwent radical resection for HCC at Zhejiang Cancer Hospital between January 2016 and December 2020 (Figure 1A).

The inclusion criteria for the study were as follows:

1. Patients who did not receive neoadjuvant therapy such as chemotherapy, radiotherapy, or interventional treatment;
2. Patients who underwent enhanced CT examination within 1 month before surgery;
3. Complete clinical history and standardized and systematic follow-up information at least 6 months of follow-up;

The exclusion criteria were as follows:

1. Patients with a previous history of other malignancies;
2. Patients with severe vascular invasion, portal vein aneurysm emboli, or extrahepatic metastases identified on preoperative imaging;
3. Patients who died within 30 days after surgery due to surgical complications;

The study was conducted following the principles of the Declaration of Helsinki. The Ethics Committee of Zhejiang Cancer Hospital approved this retrospective study, and the requirement for informed patient consent was waived (IRB-2022-503).

3 MODEL CONSTRUCTION

4 RESULTS

4.2 Feature analysis

To identify features associated with HCC early recurrence, we conducted LASSO analysis with cross-validation and found five features, including two radiomics and three pathomics features (Figure 2). We evaluated the robustness of these features by comparing their differences between the early recurrence and without early recurrence groups using the Wilcoxon rank sum test (Figure S1). Ultimately, we arrived at the following expression for the combination of features:

$$$$ {\displaystyle \begin{array}{c}\mathrm{Radiomics}\ \mathrm{Score}=0.116582{3}^{\ast}\mathrm{Glcm}\_\mathrm{InverseDifferenceMomentNormalized}\hfill \\ {}\kern8em +0.170739{0}^{\ast}\mathrm{Shape}\_\mathrm{MajorAxisLength}.\hfill \end{array}} $$$$

$$$$ \mathrm{Pathomics}\ \mathrm{Score}=-0.00347133{8}^{\ast}\mathrm{Granularity}\_1\_\mathrm{H}\&\mathrm{E}. $$$$

$$$$ +0.13619042{7}^{\ast}\mathrm{Granularity}\_4\_\mathrm{Eosin}. $$$$

$$$$ -0.28394639{2}^{\ast}\mathrm{Intensity}\_\mathrm{UpperQuartileIntensity}\_\mathrm{Eosin}. $$$$

4.3 Model performance

This study utilized nine characteristics associated with early recurrence to develop four prediction models (CRP, CRp, CrP, and rCP). A heat map (Figure 3) presented the degree of association between these seven variables. Evaluation of these models was done using SVM, LR, GNB, and KNN modeling, followed by 5-fold cross-model validation and confusion matrix analysis. In addition, Table S4 shows the change in prediction performance over five cross-validations. The CRP multimodal model outperformed the bimodal models for predicting early postoperative recurrence, with the AUC ranging from 0.743 to 0.863 (Figure 4). The AUC ranged from 0.729 to 0.813, 0.741 to 0.837, and 0.692 to 0.811 for the CRp, CrP, and cRP models, respectively (Figure 4). This suggests that omitting one type of feature set can slightly or moderately impair prediction performance.

Notably, the SVM (best parameters: “kernel type”: “linear,” “regularization parameter (C)”: 1, “gamma”: 1^e-05) showed the best performance in the multimodal prediction model (AUC: 0.863, ACC: 0.784, sensitivity: 0.731, specificity: 0.826) (Figure 4, Table 2). In addition, radiomics score and pathomics score effectively stratified HCC outcomes, with higher scores associated with higher overall recurrence risk and OS (Figure 4E–H).

TABLE 2. The average AUC, ACC, sensitivity and specificity by 16 models after 5-Fold Cross-Validation for Identifying early recurrence.

Models	Classifiers	AUC	ACC	Sensitivity	Specificity
CRP	SVM	0.863	0.784	0.731	0.826
	LR	0.840	0.766	0.731	0.792
	GNB	0.859	0.774	0.691	0.841
	KNN	0.743	0.711	0.651	0.761
CRp	SVM	0.813	0.756	0.731	0.774
	LR	0.810	0.710	0.751	0.671
	GNB	0.807	0.691	0.669	0.705
	KNN	0.729	0.719	0.691	0.741
CrP	SVM	0.827	0.719	0.691	0.742
	LR	0.831	0.728	0.711	0.741
	GNB	0.837	0.709	0.631	0.777
	KNN	0.741	0.627	0.633	0.624
cRP	SVM	0.802	0.691	0.589	0.776
	LR	0.811	0.710	0.651	0.756
	GNB	0.802	0.681	0.504	0.824
	KNN	0.692	0.635	0.524	0.723

• Abbreviations: ACC, accuracy; AUC, area under the curve; CRP, c-reactive protein; CrP, Clinical-Pathomic model; CRp, Clinical-Radiomic model; CRP, Clinical-Radiomic-Pathomic model; cRP, Radiomic-Pathomic model; GNB, Gaussian Naive Bayes; KNN, K neighbor algorithm; LR, logistic regression; SVM, Support Vector Machine.

4.4 Prognostic performance of the comprehensive nomogram

For ease of clinical application, based on the entire dataset, we constructed nomograms using the best clinical features, radiomics, and pathomics scores to represent the model results visually. Furthermore, the AUC and calibration curves of the composite nomogram indicated high accuracy and reliability of our model predictions, making it a valid tool for clinical practice in predicting the risk of early recurrence (Figure 5, Figure S2, and Table S5).

5 DISCUSSION

In this study, a multimodal model combining clinical radiomic pathomics features was constructed and validated by five-fold cross-validation to predict the risk of early recurrence of HCC after radical resection. The multimodal model features proved to have better performance than the bimodal model. In addition, the SVM constructed based on multimodal features outperformed other machine learning algorithms in predicting early recurrence. The machine learning model developed in this study provides essential support for more accurate personalized treatment plans and helps to optimize treatment decisions further, reduce the postoperative recurrence rate, and improve patients' quality of life.

In recent years, radiomics, as an emerging technology, can provide a large amount of information on morphological features, texture features, etc., which reflect the complexity of tumors and predict the growth and metastasis trends of tumors, and can be used to predict the risk of tumor recurrence.^{35, 36} However, there are still some limitations in using radiomics technology alone, such as inconsistency and uncertainty in feature extraction.³⁷ To overcome these issues, researchers have combined radiomics technology with clinical features to obtain more comprehensive tumor information and improve predictive performance. Several recent studies have demonstrated that combining radiomics and clinical features can improve the accuracy of recurrence prediction in liver and pancreatic cancers and obtain better predictive performance than alone.^{11, 12, 38} Similarly, in our study, we found that both etiology and increased AFP can independently increase the risk of HCC recurrence, and when we combined radiomics features with clinical features, we found a significant improvement in predictive performance, consistent with previous research results.^{39, 40} It is worth noting that in our previous similar study, etiology, AFP, and tumor size were selected for clinical features, and there was a difference in the radiomic features selected for the two studies.⁴¹ This difference may be related to factors such as the number and composition of patients and the definition of recurrence.

With the continuous development of WSI technology, pathology-based analysis of tissue pathology slides has been widely used in tumors.^{42, 43} In contrast to the traditional qualitative diagnosis of pathology, which relies on the subjective experience of pathologists, histopathology transforms pathological images into high-fidelity, high-throughput datasets through digital technology. This innovative approach enables quantitative analysis of pathology, resulting in a series of quantitative metrics, such as image quality features, image intensity features, image colocalization features, etc., which make pathological diagnosis more objective and reliable. Recent studies have shown that machine learning algorithms can accurately predict patients' survival and tumor grading by analyzing bladder cancer WSI.⁴⁴ In addition, pathological proteomics has also demonstrated exemplary performance in predicting ovarian and liver cancer recurrence.^{19, 45} In our study, various machine learning algorithms were modeled and compared using clinical and pathological features, and a highly accurate and reliable clinical pathological proteomics model was ultimately developed, further validating its reliability in predicting liver cancer recurrence. It is important to note that the information provided by pathomics digs deep into the microscopic level of the tumor cells and their subcellular structures, which includes an exhaustive description of the nuclear morphology, chromatin distribution, organelle organization, and cytoplasmic structure, which is a direct reflection of the tumor's malignancy and therapeutic sensitivity. In contrast, the information presented by radiomics tends to reveal the macroscopic features of the tumor tissue. It captures the tumor's overall shape, size, location, and relative relationship with surrounding tissues through medical imaging techniques. This information emphasizes the general characteristics of the tumor tissue in terms of intensity distribution, density differences, and spatial relationships.^{20, 21} Therefore, combining radiomics and pathomics features can comprehensively capture both macroscopic and microstructural features of tumors and different dimensions of information at the tissue and cellular levels, which can help to understand the nature of the disease more accurately and guide therapeutic decisions.

Our study aims to explore predictive methods for early the recurrence of HCC. However, we found no reports on the joint prediction of clinical, radiomics, and pathomics features for the early recurrence of HCC. Based on 107 patients, two radiomics features, three pathomics features, and two clinical features were finally selected. We compare multiple learning algorithms and different types of features to determine the most favorable method for performance, ultimately developing a multimodal CRP-SVM model with predictive ability. The main reasons for the model's excellent predictive ability are as follows: (a) data sources: clinical, radiomics, and pathomics features from different dimensions can be used to evaluate tumors and predict disease prognosis and treatment outcomes more effectively⁴⁶; (b) data selection: the multi-factor regression algorithm and LASSO algorithm can effectively reduce data overfitting, filter out a large number of features that have little impact on repetition, and improve prediction accuracy; (c) modeling method: SVM has the advantages of dealing with nonlinear problems in high-dimensional space, processing small sample data, adapting to different types of datasets, avoiding local minimum value problems, and strong generalization ability²⁹; (d) validation method: although our study is a small sample study, we implemented 5-fold cross-validation to balance interclass deviation.⁴⁷

Finally, we constructed a more readable and usable visual nomogram based on the radiomics score, pathomics score, and clinical features to provide decision support for clinicians to provide personalized treatment. First, by analyzing each patient's comprehensive radiomic, pathomic, and clinical features, physicians can better understand the patient's probability of disease risk and develop a personalized treatment plan. Second, predictive assessment using this model can more accurately predict a patient's risk of early recurrence and develop personalized long-term follow-up plans and targeted monitoring. However, it is worth noting that this customized approach to treatment and prognostic assessment also faces some challenges. First, the approach may require additional technology and resources, including advanced imaging techniques and pathology analysis. Second, biological differences in individual patients may lead to some model limitations. Therefore, when implementing this approach, we need to weigh its potential benefits against cost, feasibility, and applicability.

Although we have identified some promising findings, our study has some limitations. First, our dataset was retrospectively collected from a single institution with limited patient numbers. Although we implemented 5-fold cross-validation in our study to mitigate this limitation, future research should use larger, multi-center prospective datasets. Second, many factors can affect the reproducibility of features, such as segmentation methods and radiomics feature extraction software. In this study, we took measures such as having two radiologists consistently delineate the CT image's ROI to improve the reproducibility of features. However, these measures can only partially eliminate the problem of reproducibility. Third, pathomics features were sourced from tissue biopsies, which may have tissue heterogeneity and sampling bias.

6 CONCLUSION

A multimodal model, which combines clinical, radiomics, and pathomics features, was developed to predict early recurrence in HCC patients who underwent radical surgery, demonstrating high predictive performance. This model could provide valuable references for personalized treatment and monitoring plans for HCC patients.

AUTHOR CONTRIBUTIONS

Qu Xie: Conceptualization (equal); data curation (equal); formal analysis (equal); validation (equal); visualization (equal); writing – original draft (equal). Zeyin Zhao: Conceptualization (equal); software (equal). Yanzhen Yang: Formal analysis (equal); methodology (equal). Xiaohong Wang: Methodology (equal); project administration (equal). Wei Wu: Investigation (equal). Haitao Jiang: Conceptualization (equal); investigation (equal). Weiyuan Hao: Project administration (equal). Ruizi Peng: Conceptualization (equal); project administration (equal). Cong Luo: Conceptualization (equal); data curation (equal); formal analysis (equal); funding acquisition (equal).

CONFLICT OF INTEREST STATEMENT

The authors declare that the research was conducted without any commercial or financial relationships that could be construed as a potential conflict of interest.

ETHICS STATEMENT

Patients were recruited following the Helsinki Declaration. The Ethics Committee of Zhejiang Cancer Hospital approved this retrospective study, and the requirement for informed patient consent was waived (IRB-2022-503).